Sentinel-2 Land Cover Segmentation Dataset

This dataset supports the study "Transformer-Based Land Cover Classification and Multi-Temporal Change Detection in Sandu District, The Gambia: A Validated Deep Learning Pipeline for Agricultural Land Monitoring." It contains Sentinel-2 multispectral satellite imagery composites (2018, 2020, 2022, 2024, and 2025), ESA WorldCover-derived land cover labels for the 2020 reference year, extracted training/validation/test image patches used to fine-tune a SegFormer-B0 semantic segmentation model, and the resulting trained model weights. The study area is Sandu district, Upper River Region, The Gambia (344 km²). Imagery was retrieved via Google Earth Engine at 10 m resolution across six spectral/index bands (B2, B3, B4, B8, B11, B12, plus NDVI, NDWI, and NDBI). Labels follow a six-class land cover taxonomy (Tree cover, Shrubland, Grassland, Cropland, Built-up, Bare land), derived from ESA WorldCover and independently validated via manual point-based inspection (n=240, 94.2% agreement).